Periodic Locking and Topic Retention in Long Dialogues Between Two Small Language Model Agents

Abstract

Objective: To examine the joint evolution of textual repetition, retention of opening-topic information, and generation costs in long dialogues between two agents using a small language model with fixed weights and no external task feedback, and to test the effects of simple prompts and branching interventions on repetitive states. Methods: Qwen3.5-0.8B generated alternating dialogues under five fictional Chinese opening scenarios and two system-prompt conditions. The analysis included 621 complete conversations of 200 messages each, totaling 124,200 messages; 620 conversations formed 310 matched pairs. Trajectories were characterized using exact-text periodicity, character-bigram similarity, lexical novelty, Chinese semantic embeddings, grouped cross-validation, and paired resampling. Ten frozen history prefixes were also continued under five conditions and two new seeds, yielding 100 branches and 1,200 additional messages. Results: The final 40 messages of 610 conversations (98.23%) exhibited a strict exact-text period of at most 10 messages; periods 1 and 2 accounted for 299 and 280 conversations, respectively. Same-speaker exact repetition reached 94.60% in the final 20 messages, while opening-category decoding accuracy remained 71.82% using the final two messages. Adding a topic-anchoring instruction produced a paired difference of 1.13 percentage points in overall same-speaker repetition, with a 95% confidence interval of −2.21 to 4.49 percentage points under replicate-block resampling. Logistic regression using features from the first 20 messages achieved an AUC of 0.720 for predicting confirmation of near-repetition at or before message 80; more complex models showed no consistent advantage. A stopping rule requiring 20 consecutive near-repetitions would retrospectively have avoided 63.57% of messages and 85.77% of recorded generation time. In the branch experiment, the text-change fraction across all 12 continuation messages was 7.50% under the recommended sampling configuration and zero under the original configuration and the other three interventions. The exact cluster test yielded p=0.25 and the Benjamini–Hochberg-adjusted q value for the primary comparisons was 1.00; sustained change across the final six messages occurred in only 1/20 branches under the recommended configuration. Conclusions: Under the model, opening scenarios, template protocol, and sampling settings examined here, periodic locking can coexist with retention of opening-topic information. Evaluation of long dialogues should separately measure structural repetition, topic identifiability, and informational usefulness, while using controlled interventions to test the mutability of repetitive states.